《金融投资学》教学资源（英文文献）Conditional Methods in event study.pdf

The Review of Financial Studies /v 10 n 1 1997 Event studies are widely used to study the information content of corporate events.Such studies typically have two purposes:(1)to test for the existence of an "information effect"(i.e.,the impact of an event on the announcing firm's value)and to estimate its magnitude,and (2)to identify factors that explain changes in firm value on the event date. To test for the existence of an information effect,empirical finance has primarily employed the technique developed in Fama,Fisher, Jensen,and Roll (1969)(referred to as FFJR hereafter).FFJR suggest that if an event has an information effect,there should be a nonzero stock-price reaction on the event date.Thus,inference is based on the statistical significance of the average announcement effect for a sample of firms announcing the event in question.The FFJR test is usually followed by a linear regression of announcement effects on a set of firm-specific factors to identify those factors that explain the cross-section of announcement effects.Most event studies in the applied literature have been based on the above methods.2 However,recent literature on conditional event-study methods [Acharya (1988,1993),Eckbo,Maksimovic,and Williams (1990)]ar- gues that the traditional methods are misspecified in a rational ex- pectations context.Briefly,the argument is that corporate events are voluntary choices of firms and are typically initiated when firms come to possess information not fully known to markets.The unexpected portion of such information should determine the stock-price reaction to the event. When events are modeled in this manner within simple equilib- rium settings,the resulting specifications are typically nonlinear cross- sectional regressions3 that bear little resemblance to the simple mod- els conventionally used in event studies.Hence,it has been suggested that the conventional methods are misspecified and lead to unreliable inferences,implying that such methods should not be used in prac- tice.More generally,this debate does raise the important issue that though the standard event-study procedures have been widely used in empirical work,little is understood about their consistency and power Announcement effect (or abnormal return)denotes the excess of the actual event-date stock return over the unconditional expected return for the stock.The latter is usually estimated via the market model,calibrated on pre-event data [see Brown and Warner(1985)for a more complete discussionl. 2A partial list of applications includes studies of (1)equity and debt issues [Asquith and Mullins (1986),Eckbo (1986)1,(2)timing of equiry issues [Korajczyk,Lucas,and McDonald (1991)1,(3) takeovers [Asquith,Bruner,and Mullins (1983),(4)dividends [Bajaj and Vijh (1990),and (5)stock repurchases [Vermaelen (1984). 3The nonlinearity stems from the endogeneity of events.Endogeneiry truncates the statistical dis- tribution of announcement effects. 2

The Review of Financial Studies / v 10 n 1 1997 Event studies are widely used to study the information content of corporate events. Such studies typically have two purposes: (1) to test for the existence of an “information effect” (i.e., the impact of an event on the announcing firm’s value) and to estimate its magnitude, and (2) to identify factors that explain changes in firm value on the event date. To test for the existence of an information effect, empirical finance has primarily employed the technique developed in Fama, Fisher, Jensen, and Roll (1969) (referred to as FFJR hereafter). FFJR suggest that if an event has an information effect, there should be a nonzero stock-price reaction on the event date. Thus, inference is based on the statistical significance of the average announcement effect1 for a sample of firms announcing the event in question. The FFJR test is usually followed by a linear regression of announcement effects on a set of firm-specific factors to identify those factors that explain the cross-section of announcement effects. Most event studies in the applied literature have been based on the above methods.2 However, recent literature on conditional event-study methods [Acharya (1988, 1993), Eckbo, Maksimovic, and Williams (1990)] argues that the traditional methods are misspecified in a rational expectations context. Briefly, the argument is that corporate events are voluntary choices of firms and are typically initiated when firms come to possess information not fully known to markets. The unexpected portion of such information should determine the stock-price reaction to the event. When events are modeled in this manner within simple equilibrium settings, the resulting specifications are typically nonlinear crosssectional regressions3 that bear little resemblance to the simple models conventionally used in event studies. Hence, it has been suggested that the conventional methods are misspecified and lead to unreliable inferences, implying that such methods should not be used in practice. More generally, this debate does raise the important issue that though the standard event-study procedures have been widely used in empirical work, little is understood about their consistency and power 1 Announcement effect (or abnormal return) denotes the excess of the actual event-date stock return over the unconditional expected return for the stock. The latter is usually estimated via the market model, calibrated on pre-event data [see Brown and Warner (1985) for a more complete discussion]. 2 A partial list of applications includes studies of (1) equity and debt issues [Asquith and Mullins (1986), Eckbo (1986)], (2) timing of equity issues [Korajczyk, Lucas, and McDonald (1991)], (3) takeovers [Asquith, Bruner, and Mullins (1983)], (4) dividends [Bajaj and Vijh (1990)], and (5) stock repurchases [Vermaelen (1984)]. 3 The nonlinearity stems from the endogeneity of events. Endogeneity truncates the statistical distribution of announcement effects. 2

Conditional Metbods in Event Studies in rational expectations settings,such as those underlying conditional methods. This article has three purposes.First,we present a simple exposition of conditional methods that focuses on their economic content.We show that all conditional models have essentially the same economic intuition,and derive all received models within a common framework that reflects this perspective.This synthesis reconciles different speci- fications proposed in the literature,clarifies their shared intuition,and suggests how one might choose between or extend such models in practice. Our second point is that while traditional event-study techniques are indeed misspecified in the conditional methods context,they still lead to valid inferences,under certain statistical conditions described in this article.Specifically,even when event-study data are generated exactly as per conditional models of the sort introduced by Acharya (1988),(1)the FFJR procedure remains a well-specified test for de- tecting the existence of information effects;and (2)the conventional cross-sectional procedure yields parameter estimates proportional to the true conditional model parameters,under the conditions men- tioned before.The proportionality factor has a simple interpretation in terms of the informational parameters of the event.These results provide,for the first time,an equilibrium justification for the proce- dures conventionally used to conduct event studies. Finally,if both traditional and conditional methods lead to equiv- alent inferences,how does one choose between the two in practice? Working in the context of the conditional model proposed by Acharya (1988),we develop simulation evidence on this issue.Our evidence suggests that one's choice would depend mainly on whether one has, besides the usual event-study data,an additional sample of"nonevent" firms,that is,firms that were partially anticipated to announce but did not announce the event in question.If such nonevent data are avail- able,conditional methods are powerful means of conducting event studies and may be implemented effectively using a simple"two-step" estimator.Absent nonevent data,conditional methods appear to offer little value relative to traditional procedures. This article is organized along the above lines.Section 1 presents conditional methods for event studies.Section 2 presents and dis- cusses the main analytic results,regarding the equivalence of infer- ences via conditional and traditional event-study methods.Section 3 motivates the question of choosing between the two approaches,and Section 4 outlines the structure of the simulations conducted to ad- dress this question.Simulation results are presented in Section 5,and Section 6 offers concluding comments. 3

Conditional Methods in Event Studies in rational expectations settings, such as those underlying conditional methods. This article has three purposes. First, we present a simple exposition of conditional methods that focuses on their economic content. We show that all conditional models have essentially the same economic intuition, and derive all received models within a common framework that reflects this perspective. This synthesis reconciles different speci- fications proposed in the literature, clarifies their shared intuition, and suggests how one might choose between or extend such models in practice. Our second point is that while traditional event-study techniques are indeed misspecified in the conditional methods context, they still lead to valid inferences, under certain statistical conditions described in this article. Specifically, even when event-study data are generated exactly as per conditional models of the sort introduced by Acharya (1988), (1) the FFJR procedure remains a well-specified test for detecting the existence of information effects; and (2) the conventional cross-sectional procedure yields parameter estimates proportional to the true conditional model parameters, under the conditions mentioned before. The proportionality factor has a simple interpretation in terms of the informational parameters of the event. These results provide, for the first time, an equilibrium justification for the procedures conventionally used to conduct event studies. Finally, if both traditional and conditional methods lead to equivalent inferences, how does one choose between the two in practice? Working in the context of the conditional model proposed by Acharya (1988), we develop simulation evidence on this issue. Our evidence suggests that one’s choice would depend mainly on whether one has, besides the usual event-study data, an additional sample of “nonevent” firms, that is, firms that were partially anticipated to announce but did not announce the event in question. If such nonevent data are available, conditional methods are powerful means of conducting event studies and may be implemented effectively using a simple “two-step” estimator. Absent nonevent data, conditional methods appear to offer little value relative to traditional procedures. This article is organized along the above lines. Section 1 presents conditional methods for event studies. Section 2 presents and discusses the main analytic results, regarding the equivalence of inferences via conditional and traditional event-study methods. Section 3 motivates the question of choosing between the two approaches, and Section 4 outlines the structure of the simulations conducted to address this question. Simulation results are presented in Section 5, and Section 6 offers concluding comments. 3

Tbe Review of Financial Studies/v 10 n 1 1997 1.On Conditional Methods Section 1 develops conditional models for event studies.The main point made here is that all conditional models have essentially the same economic intuition:they relate announcement effects to the un- expected information revealed in events.While the notion of relating announcement effects to unexpected information is not new,we show here that it is the common theme that underlies all conditional models We demonstrate that all received models may be derived in terms of this framework,and that the models differ only because they make dif- ferent assumptions about the information structure underlying events. The exposition proceeds as follows.Section 1.1 opens with a dis- cussion of the intuition underlying conditional methods.Section 1.2 discusses alternative ways of modeling the information structure in events,and Sections 1.3 through 1.5 develop econometric models for announcement effects for each of these information structures. 1.1 Intuition underlying conditional methods To begin,note the potential dichotomy between the fact of an event and the information it reveals.For instance.the event "takeover"is plausibly less surprising for a bidder with announced acquisition pro- grams than for a bidder with no history of acquisitions.Similarly,the event"dividend increase"is less surprising for a firm with an unusually good spell of earnings than for a firm with flat or declining earnings. Thus,a given event may convey less information for some firms and more for others.Further,it should be the unexpected information revealed in events that causes the stock-price changes around event dates.4 This discussion suggests the following empirical procedure for car- rying out event studies:(1)estimate for each firm the unexpected information that the event reveals:(2)compute the cross-sectional correlation between information and abnormal return and test for its significance.A nonzero correlation would indicate that abnormal return is systematically related to information revealed in the event (i.e.,there exists an information effect).Conversely,zero correlation implies lack of an information effect.This intuition underlies every conditional specification analyzed in this article. Central to the conditional paradigm is the notion of"information revealed in events.Next,we discuss how this might be modeled in the event-study context. Malatesta and Thompson (1985),Thompson (1985),and Chaplinsky and Hansen (1993)also recognize the role of partial anticipation of events and examine its implications for event studies based on FFJR-style procedures

The Review of Financial Studies / v 10 n 1 1997 1. On Conditional Methods Section 1 develops conditional models for event studies. The main point made here is that all conditional models have essentially the same economic intuition: they relate announcement effects to the unexpected information revealed in events. While the notion of relating announcement effects to unexpected information is not new, we show here that it is the common theme that underlies all conditional models. We demonstrate that all received models may be derived in terms of this framework, and that the models differ only because they make different assumptions about the information structure underlying events. The exposition proceeds as follows. Section 1.1 opens with a discussion of the intuition underlying conditional methods. Section 1.2 discusses alternative ways of modeling the information structure in events, and Sections 1.3 through 1.5 develop econometric models for announcement effects for each of these information structures. 1.1 Intuition underlying conditional methods To begin, note the potential dichotomy between the fact of an event and the information it reveals. For instance, the event “takeover” is plausibly less surprising for a bidder with announced acquisition programs than for a bidder with no history of acquisitions. Similarly, the event “dividend increase” is less surprising for a firm with an unusually good spell of earnings than for a firm with flat or declining earnings. Thus, a given event may convey less information for some firms and more for others. Further, it should be the unexpected information revealed in events that causes the stock-price changes around event dates.4 This discussion suggests the following empirical procedure for carrying out event studies: (1) estimate for each firm the unexpected information that the event reveals; (2) compute the cross-sectional correlation between information and abnormal return and test for its significance. A nonzero correlation would indicate that abnormal return is systematically related to information revealed in the event (i.e., there exists an information effect). Conversely, zero correlation implies lack of an information effect. This intuition underlies every conditional specification analyzed in this article. Central to the conditional paradigm is the notion of “information revealed in events.” Next, we discuss how this might be modeled in the event-study context. 4 Malatesta and Thompson (1985), Thompson (1985), and Chaplinsky and Hansen (1993) also recognize the role of partial anticipation of events and examine its implications for event studies based on FFJR-style procedures. 4

Conditional Metbods in Event Studies 1.2 Specifying the information structure As argued before,events reveal the information that their announce- ment is conditioned on.Suppose that this information consists of a variable ti,which arrives at firm i on an information arrival date. Information ti is subsequently revealed to markets,via the event,on an event date. What do markets learn from the revelation of t?Clearly,this de- pends on what markets knew,prior to the actual event date,about the arrival of information t;at firm i.Here,we allow for three possibilities: Assumption 1.Markets know,prior to the event,that the event-related information t;bas arrived at firm i (but not its exact content). Assumption 2.Markets do not know,prior to the event,that infor- mation ti bas arrived at firm i. Assumption 3.Markets assess a probability p (0,1)that informa- tion ti bas arrived at firm i. Under Assumption 1,information arrival is common knowledge prior to the event;under Assumption 2,markets do not know about information arrival prior to the event-date.Finally,Assumption 3 is the encompassing case that permits markets to make probabilistic assessments about information arrival.5 Each assumption leads to a different econometric specification for announcement effects,as we show below. For expositional ease and because previous work has been based on Assumptions 1 and 2,we first develop the methodology under Assumptions 1 and 2,in Sections 1.3 and 1.4.We then present an encompassing specification,based on Assumption 3,in Section 1.5. 1.3 Model I:information arrival known prior to event We begin by making Assumption 1:that markets know,prior to the event,that the event-related information t;has arrived at firm i.In general,this leads markets to form expectations about ti.Suppose these expectations are given by E1(t)=Ex1= (1) =1 5 The following example illustrates the distinction between the three assumptions.Consider the event "takeover"and suppose takeovers occur if and only if the acquirer-bidder synergy (T)is positive.Assumption 1 implies that markets always know,prior to each takeover announcement, that the bidder had identified the target in question.The only uncertainty is whether t is positive or not.In contrast,Assumption 2 implies that markets do not know,prior to each announcement, that the target had been identified.Under Assumption 3,markets assign probabiliry ps(0,1)that the target had been identified. 5

Conditional Methods in Event Studies 1.2 Specifying the information structure As argued before, events reveal the information that their announcement is conditioned on. Suppose that this information consists of a variable τi, which arrives at firm i on an information arrival date. Information τi is subsequently revealed to markets, via the event, on an event date. What do markets learn from the revelation of τi? Clearly, this depends on what markets knew, prior to the actual event date, about the arrival of information τi at firm i. Here, we allow for three possibilities: Assumption 1. Markets know, prior to the event, that the event-related information τi has arrived at firm i (but not its exact content). Assumption 2. Markets do not know, prior to the event, that information τi has arrived at firm i. Assumption 3. Markets assess a probability p ∈ (0, 1) that information τi has arrived at firm i. Under Assumption 1, information arrival is common knowledge prior to the event; under Assumption 2, markets do not know about information arrival prior to the event-date. Finally, Assumption 3 is the encompassing case that permits markets to make probabilistic assessments about information arrival.5 Each assumption leads to a different econometric specification for announcement effects, as we show below. For expositional ease and because previous work has been based on Assumptions 1 and 2, we first develop the methodology under Assumptions 1 and 2, in Sections 1.3 and 1.4. We then present an encompassing specification, based on Assumption 3, in Section 1.5. 1.3 Model I: information arrival known prior to event We begin by making Assumption 1: that markets know, prior to the event, that the event-related information τi has arrived at firm i. In general, this leads markets to form expectations about τi. Suppose these expectations are given by E−1(τi) = θ0 xi = Xn j=1 θjxij (1) 5 The following example illustrates the distinction between the three assumptions. Consider the event “takeover” and suppose takeovers occur if and only if the acquirer-bidder synergy (τ ) is positive. Assumption 1 implies that markets always know, prior to each takeover announcement, that the bidder had identified the target in question. The only uncertainty is whether τ is positive or not. In contrast, Assumption 2 implies that markets do not know, prior to each announcement, that the target had been identified. Under Assumption 3, markets assign probability p ε (0, 1) that the target had been identified. 5